A zero-watermark generation method and system of a lightweight three-dimensional lung model

Abstract

The application discloses a zero watermark generation method and system of a lightweight three-dimensional lung model, comprising the following steps: using a 3D U-Net network to perform anatomical subregion segmentation on a three-dimensional lung simplified model to extract subregion style features; checking and screening the subregion style features according to cosine similarity to obtain independent sub-feature vectors; performing hierarchical extraction on the independent sub-feature vectors to obtain multi-scale features; performing cross-scale fusion on the multi-scale features to obtain global style features; inputting the global style features into a 3D MobileViT lightweight network to obtain key style features, performing PCA feature dimension reduction on the key style features to obtain lightweight features; splicing the lightweight features and copyright information to obtain copyright splicing features, and generating a zero watermark sequence of the copyright splicing features by using a hash operation; and the method realizes accurate protection of key copyright information of the three-dimensional lung model and significantly improves the robustness and efficiency of the algorithm.

Description

Technical Field

[0001] This invention belongs to the field of watermark data processing technology, specifically relating to a method and system for generating zero watermarks for lightweight three-dimensional lung models. Background Technology

[0002] With the continuous development of the integration of 3D reconstruction technology and medical imaging, the importance of 3D lung models in the medical field is becoming increasingly prominent. However, copyright infringement of 3D models is on the rise, mainly manifested in local structural alterations, feature distortion caused by format conversion, and asymmetric cropping, which poses a significant challenge to the copyright protection of 3D lung models.

[0003] Most existing zero-watermarking technologies are designed for two-dimensional images, and they have many significant limitations when transferred to three-dimensional models. First, existing technologies suffer from global feature dependency, making them difficult to effectively handle local tampering. For example, current 3D zero-watermarking technologies often use global features, failing to fully consider the differences in local anatomical features of different organs. Second, existing technologies are insufficient in scale adaptability, resulting in limited robustness. Three-dimensional models possess multi-scale features at micro, meso, and macro levels, and these features are closely correlated. However, existing zero-watermarking technologies follow two-dimensional multi-scale strategies, which cannot match the scale logic of three-dimensional structures.

[0004] In summary, existing technologies have many shortcomings in protecting the copyright of 3D lung models, making it difficult to meet the high requirements of copyright protection in practical applications. Therefore, there is an urgent need for a zero-watermark technology for 3D lung models that can effectively resist local tampering, has good scale adaptability, and high extraction efficiency, in order to better protect the copyright of 3D lung models. Summary of the Invention

[0005] This invention provides a lightweight three-dimensional lung model zero-watermark generation method and system, which realizes accurate protection of key copyright information of the three-dimensional lung model, significantly improves the robustness and efficiency of the algorithm, reduces the risk of verification failure, and has good adaptability and scalability.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] The first aspect of this invention provides a method for generating a zero-watermark lightweight three-dimensional lung model, comprising:

[0008] Adaptive sparse voxel sampling is performed on the original 3D lung model to obtain a simplified 3D lung model;

[0009] A 3D U-Net network was used to segment anatomical subregions of a simplified 3D lung model and extract subregion style features; the cosine similarity between the subregion style features and the anatomical semantic labels was calculated. Based on cosine similarity The sub-region style features are verified and filtered to obtain independent sub-feature vectors;

[0010] Multi-scale features are obtained by hierarchically extracting independent sub-feature vectors; global style features are obtained by cross-scale fusion of multi-scale features.

[0011] Global style features are input into the 3D MobileViT lightweight network to obtain key style features, and PCA feature dimensionality reduction is performed on the key style features to obtain lightweight features.

[0012] The lightweight features and copyright information are vector-concatenated to obtain the copyright concatenation features, and a zero-watermark sequence of the copyright concatenation features is generated using hash operations.

[0013] Furthermore, adaptive sparse voxel sampling is performed on the original 3D lung model to obtain a simplified 3D lung model, specifically including:

[0014] Calculate the voxel gray variance of the original three-dimensional lung model, divide the original three-dimensional lung model into lung voxel dense region and lung voxel sparse region according to the voxel gray variance, and set the sampling rate of lung voxel dense region and lung voxel sparse region.

[0015] The first voxel sampling feature is obtained by randomly sampling the sparse voxel region of the lung according to the set sampling rate, and the second voxel sampling feature is obtained by grid sampling the dense voxel region of the lung according to the set sampling rate.

[0016] A simplified three-dimensional lung model was obtained by constructing the first voxel sampling features and the second voxel sampling features.

[0017] Furthermore, a 3D U-Net network is used to segment anatomical subregions of a simplified 3D lung model and extract subregion style features, specifically including:

[0018] The 3D U-Net network includes an input layer, an encoder, a decoder, and an output layer connected in sequence; the encoder contains four 3D convolutional modules. The decoder contains four layers of deconvolution modules. 3D convolution module in encoder With the deconvolution module in the decoder Through residual connections; where... This is the sequence number of the 3D convolution module. This is the sequence number of the deconvolution module. ;

[0019] The simplified 3D lung model is input into the 3D U-Net network through the input layer. The encoder upsamples the simplified 3D lung model to obtain 3D spatial information, and the decoder downsamples the 3D spatial information. The output layer outputs the sub-region segmentation features and sub-region types.

[0020] Based on the subregion type, lung lobe subregion features, tracheal tree subregion features, and lesion subregion features are extracted from the subregion segmentation features, respectively.

[0021] The features of the lung lobe subregion include the mean gray value of the lung lobe, the variance of the lung lobe gray value, the skewness of the lung lobe gray value distribution, and the kurtosis of the lung lobe gray value distribution; the features of the tracheal tree subregion include the number of tracheal branches, the angle between tracheal branches, and the length of tracheal branches; the features of the lesion subregion include the mean curvature of the lesion surface and the Gaussian curvature of the lesion surface.

[0022] The features of the lung lobe subregion, tracheal tree subregion, and lesion subregion are used as subregion style features.

[0023] Furthermore, the cosine similarity between sub-region style features and anatomical semantic labels is calculated. Specifically, it includes:

[0024]

[0025] In the formula, Cosine similarity between sub-region style features and anatomical semantic labels; For sub-region style feature vectors; This is the standard feature vector for dissecting semantic labels; and Subregion style feature vector and standard eigenvectors The Euclidean norm.

[0026] Furthermore, based on cosine similarity The sub-region style features are validated and filtered to obtain independent sub-feature vectors, specifically including:

[0027] If cosine similarity <Set a similarity threshold A to determine abnormal sub-region style feature extraction, and re-use the 3DU-Net network to perform anatomical sub-region segmentation and extract sub-region style features of the simplified 3D lung model;

[0028] If cosine similarity ≥ Set a similarity threshold A, and generate independent sub-feature vectors from the style features of the sub-region.

[0029] Furthermore, multi-scale features are obtained by hierarchically extracting independent sub-feature vectors, specifically including:

[0030] Voxel gray-level co-occurrence matrix features are extracted from independent sub-feature vectors according to a set microscale. These voxel gray-level co-occurrence matrix features include contrast. Correlation and energy Ene;

[0031] Based on the set mesoscale, the topological features of bronchial branches and the spatial distribution features of lung lobes are extracted from the independent sub-feature vectors.

[0032] Based on the set macroscopic scale, the spatial location features and morphological features of the lung lobes are extracted from the independent sub-feature vectors; the spatial location features of the lung lobes are the coordinates of the lung lobe center and the proportion of the lung lobe volume; the morphological features of the lung lobes are the principal axis ratio of the overall lung shape.

[0033] Furthermore, multi-scale features are fused across scales to obtain global style features, specifically including:

[0034] Z-score standardization was performed on the voxel gray-level co-occurrence matrix features, bronchial branch topological features, pulmonary lobule spatial distribution features, pulmonary lobe spatial location features, and pulmonary lobe morphological features to obtain micro-scale feature vectors, meso-scale feature vectors, and macro-scale feature vectors, respectively.

[0035] Attention weights for microscale, mesoscale, and macroscale eigenvectors are calculated using linear transformations and the Softmax function, expressed as follows:

[0036]

[0037]

[0038] In the formula, For attention weights, For scale type index identifier; It is an exponential function; It is a linear transformation function of the scale eigenvectors; It is a standardized feature vector with a defined scale; For microscale indexing; For mesoscale indexing; For macroscopic scale indexing; The weight matrix is ​​a linear transformation matrix; For bias terms;

[0039] The global style feature vector is obtained by weighting the micro-scale feature vector, meso-scale feature vector, and macro-scale feature vector according to attention weights, and then summing them. The formula is as follows:

[0040]

[0041] In the formula, This is the global style feature vector; It is the attention weight at the microscale. It is a standardized vector at the microscale; It is the attention weight at the mesoscale. It is a standardized vector at the mesoscale; It is the attention weight at the macro scale. It is a standardized vector at a macroscopic scale.

[0042] Furthermore, the 3D MobileViT lightweight network includes an input layer, three depthwise separable convolutional blocks, two local Transformer attention blocks, and an output layer connected in sequence; the depthwise separable convolutional blocks include 3D depthwise convolution, ReLU6 activation function, 3D pointwise convolution, and normalization processing layer in sequence.

[0043] Furthermore, PCA feature reduction is performed on key style features to obtain lightweight features, specifically including:

[0044] Calculate the mean vector of key style features The formula is as follows:

[0045]

[0046] In the formula, This is the mean vector of the key style features. The number of samples for key style features; It is the index of the key style feature vector; For the first Key style feature vectors;

[0047] Based on the mean vector of key style features and key style feature vectors The formula for calculating the sample covariance matrix is ​​as follows:

[0048]

[0049] In the formula, The sample covariance matrix; The number of samples for key style features;

[0050] For covariance matrix Perform eigenvalue decomposition and select the eigenvectors corresponding to the top 512 largest eigenvalues ​​to form a dimension reduction matrix;

[0051] The lightweight feature vector is obtained by multiplying the key style feature vector by the dimensionality reduction matrix, expressed by the following formula:

[0052]

[0053] In the formula, For lightweight feature vectors, For key style feature vectors, This is a dimension-reduced matrix.

[0054] Furthermore, the lightweight features and copyright information are vector-concatenated to obtain copyright concatenation features. A zero-watermark sequence of copyright concatenation features is then generated using hash operations, specifically including:

[0055]

[0056] In the formula, To construct feature vectors for copyright, For lightweight feature vectors, For identification coding, Numerical encoding for timestamps. Encoding for the lung model;

[0057] Copyright splicing feature vector conduct The zero-watermark sequence obtained by hash operation is expressed by the following formula:

[0058]

[0059] In the formula, It is a zero-watermark sequence. for Hash operation function.

[0060] A second aspect of the present invention provides an electronic terminal, characterized in that it includes a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the steps of the zero-watermark generation method of the first aspect.

[0061] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0062] This invention uses a 3D U-Net network to segment anatomical subregions of a three-dimensional lung model, extracts local style features of each subregion, and ensures the effectiveness and accuracy of the features through anatomical semantic consistency verification. This method can accurately capture copyright information of local areas such as lung lobes, tracheal trees, and lesions.

[0063] This invention constructs a multi-scale feature pyramid of "micro-meso-macro" scales and dynamically allocates the weights of each scale through a cross-scale attention mechanism, thereby strengthening scale correlation. This enables the algorithm to comprehensively capture the stylistic features of a 3D lung model and further improves the accuracy of copyright protection by adjusting feature extraction and fusion strategies for the differences in local anatomical features of different organs.

[0064] In the data preprocessing stage, this invention employs adaptive sparse voxel sampling technology, dynamically adjusting the sampling rate based on the information density of lung voxels to reduce redundant data and improve data processing efficiency. Through the use of the 3D MobileViT lightweight network and PCA dimensionality reduction technology, the dimensionality and computational cost of features are significantly reduced while maintaining feature robustness. Attached Figure Description

[0065] Figure 1 This is a flowchart of the zero-watermark generation process for a three-dimensional lung model provided in Embodiment 1 of the present invention;

[0066] Figure 2 This is a flowchart of the zero-watermark copyright verification process for a three-dimensional lung model provided in Embodiment 1 of the present invention;

[0067] Figure 3 This is a structural diagram of the 3D U-Net network provided in Embodiment 1 of the present invention. Detailed Implementation

[0068] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0069] Example 1

[0070] like Figure 1 As shown, this embodiment provides a method for generating zero-watermark lightweight three-dimensional lung models, including:

[0071] Adaptive sparse voxel sampling is performed on the original 3D lung model to obtain a simplified 3D lung model, specifically including:

[0072] Calculate the voxel gray variance of the original three-dimensional lung model, divide the original three-dimensional lung model into lung voxel dense region and lung voxel sparse region according to the voxel gray variance, and set the sampling rate of lung voxel dense region and lung voxel sparse region.

[0073] The first voxel sampling feature is obtained by randomly sampling the sparse voxel region of the lung according to the set sampling rate, and the second voxel sampling feature is obtained by grid sampling the dense voxel region of the lung according to the set sampling rate.

[0074] In this embodiment, information such as lung parenchyma and lesions in densely voxel areas of the lung (voxel gray variance > 50) is sampled at 80% (to preserve key details).

[0075] For lung cavitation, extrapulmonary background and other information in sparse voxel regions of the lung (voxel gray variance ≤ 50), a sampling rate of 20% was used (to remove redundant data).

[0076] The sampling method combines random sampling and grid sampling: grid sampling is used in information-dense areas (to ensure spatial continuity), and random sampling is used in information-sparse areas (to reduce computational load). After sampling, the amount of data is compressed to 30%-35% of the original.

[0077] A simplified three-dimensional lung model was obtained by constructing the first voxel sampling features and the second voxel sampling features.

[0078] The 3D U-Net network was used to segment anatomical subregions and extract subregion style features from a simplified 3D lung model. Specifically, this included:

[0079] like Figure 3 As shown, the 3D U-Net network includes an input layer, an encoder, a decoder, and an output layer connected in sequence; the encoder contains four 3D convolutional modules. The decoder contains four layers of deconvolution modules. 3D convolution module in encoder With the deconvolution module in the decoder Through residual connections; where... This is the sequence number of the 3D convolution module. This is the sequence number of the deconvolution module. ;

[0080] The simplified 3D lung model is input into the 3D U-Net network through the input layer. The encoder upsamples the simplified 3D lung model to obtain 3D spatial information, and the decoder downsamples the 3D spatial information. The output layer outputs the sub-region segmentation features and sub-region types.

[0081] Based on the subregion type, lung lobe subregion features, tracheal tree subregion features, and lesion subregion features are extracted from the subregion segmentation features, respectively.

[0082] The features of the lung lobe subregion include the mean gray value of the lung lobe, the variance of the gray value of the lung lobe, the skewness of the gray value distribution of the lung lobe, and the kurtosis of the gray value distribution of the lung lobe.

[0083]

[0084]

[0085]

[0086]

[0087] In the formula, This represents the average gray value of the lung lobes. The total number of voxels in a single lung lobe subregion. Indexes of individual voxels in the lung lobe subregions; For the first subregion of the lung lobe The grayscale value of a single individual; The variance of grayscale values ​​for lung lobes; The skewness of grayscale distribution in lung lobes; The kurtosis of the grayscale distribution in the lung lobe;

[0088] The tracheal tree subregion features include the number of tracheal branches, the angle between tracheal branches, and the length of tracheal branches; the angle between tracheal branches is calculated using a vector dot product, expressed as:

[0089]

[0090] In the formula, The angle between the tracheal branches. and For the tracheal tree , The direction vectors of each branch; and Direction vector and direction vector The length of the module.

[0091] The lesion subregion features (referring to abnormal areas in the 3D lung model confirmed by medical image annotation, including but not limited to pulmonary nodules (diameter ≥3mm), tumor lesions, and inflammatory infiltrates, which need to be determined based on clinical diagnostic reports or annotated datasets) include the average curvature of the lesion surface and the Gaussian curvature of the lesion surface; the expression formula is:

[0092]

[0093]

[0094] In the formula, The average curvature of the lesion surface. It is the curvature value of the three-dimensional lesion surface in the direction of the most severe curvature at a certain sampling point; It is the surface of the three-dimensional lesion at the same sampling point, and The curvature value with the gentlest bend in the direction perpendicular to the direction. The Gaussian curvature of the lesion surface;

[0095] The features of the lung lobe subregion, tracheal tree subregion, and lesion subregion are used as subregion style features.

[0096] Calculate the cosine similarity between sub-region style features and anatomical semantic labels. Specifically, it includes:

[0097]

[0098] In the formula, Cosine similarity between sub-region style features and anatomical semantic labels; For sub-region style feature vectors; This is the standard feature vector for dissecting semantic labels; and Subregion style feature vector and standard eigenvectors The Euclidean norm.

[0099] Based on cosine similarity The sub-region style features are validated and filtered to obtain independent sub-feature vectors, specifically including:

[0100] If cosine similarity <Set a similarity threshold A to determine abnormal sub-region style feature extraction, and re-use the 3DU-Net network to perform anatomical sub-region segmentation and extract sub-region style features of the simplified 3D lung model;

[0101] If cosine similarity A similarity threshold A is set, which is an independent sub-feature vector generated from the style features of the sub-region. In this embodiment, the similarity threshold A is set to 0.7.

[0102] The 3D lung model is segmented into anatomical subregions using a 3D U-Net network, and local style features of each subregion are extracted. The effectiveness and accuracy of the features are ensured by anatomical semantic consistency verification. This method can accurately capture copyright information of local areas such as lung lobes, tracheal trees, and lesions.

[0103] Multi-scale features are obtained by hierarchically extracting independent sub-feature vectors, specifically including:

[0104] Voxel gray-level co-occurrence matrix features are extracted from independent sub-feature vectors according to a set microscale. These voxel gray-level co-occurrence matrix features include contrast. Correlation and energy Ene; contrast The calculation formula is:

[0105]

[0106] In the formula, Grayscale levels; and These represent the gray values ​​of two voxels at different positions in the voxel gray-level co-occurrence matrix; In the gray-level co-occurrence matrix The probability value of the location;

[0107] Based on the set mesoscale, the topological features of bronchial branches and the spatial distribution features of lung lobes are extracted from the independent sub-feature vectors.

[0108] Based on the set macroscopic scale, the spatial location features and morphological features of the lung lobes are extracted from the independent sub-feature vectors; the spatial location features of the lung lobes are the coordinates of the lung lobe center and the proportion of the lung lobe volume; the morphological features of the lung lobes are the principal axis ratio of the overall lung shape.

[0109] Cross-scale fusion of multi-scale features to obtain global style features includes:

[0110] Z-score standardization was performed on the voxel gray-level co-occurrence matrix features, bronchial branch topological features, pulmonary lobule spatial distribution features, pulmonary lobe spatial location features, and pulmonary lobe morphological features to obtain micro-scale feature vectors, meso-scale feature vectors, and macro-scale feature vectors, respectively.

[0111] Attention weights for microscale, mesoscale, and macroscale eigenvectors are calculated using linear transformations and the Softmax function, expressed as follows:

[0112]

[0113]

[0114] In the formula, For attention weights, For scale type index identifier; It is an exponential function; It is a linear transformation function of the scale eigenvectors; It is a standardized feature vector with a defined scale; For microscale indexing; For mesoscale indexing; For macroscopic scale indexing; The weight matrix is ​​a linear transformation matrix; For bias terms;

[0115] The global style feature vector is obtained by weighting the micro-scale feature vector, meso-scale feature vector, and macro-scale feature vector according to attention weights, and then summing them. The formula is as follows:

[0116]

[0117] In the formula, This is the global style feature vector; It is the attention weight at the microscale. It is a standardized vector at the microscale; It is the attention weight at the mesoscale. It is a standardized vector at the mesoscale; It is the attention weight at the macro scale. It is a standardized vector at a macroscopic scale.

[0118] This embodiment constructs a multi-scale feature pyramid of "micro-meso-macro" scales and dynamically allocates the weights of each scale through a cross-scale attention mechanism, thereby strengthening scale correlation. This enables the algorithm to comprehensively capture the stylistic features of the 3D lung model and further improves the accuracy of copyright protection by adjusting the feature extraction and fusion strategies for the differences in local anatomical features of different organs.

[0119] The 3D MobileViT lightweight network comprises an input layer, three depthwise separable convolutional blocks, two local Transformer attention blocks, and an output layer connected in sequence. The depthwise separable convolutional blocks include, in sequence, 3D depthwise convolution, ReLU6 activation function, 3D pointwise convolution, and normalization processing layer.

[0120] Global style features are input into the 3D MobileViT lightweight network to obtain key style features. PCA feature reduction is then performed on these key style features to obtain lightweight features, specifically including:

[0121] Calculate the mean vector of key style features The formula is as follows:

[0122]

[0123] In the formula, This is the mean vector of the key style features. The number of samples for key style features; It is the index of the key style feature vector; For the first Key style feature vectors;

[0124] Based on the mean vector of key style features and key style feature vectors The formula for calculating the sample covariance matrix is ​​as follows:

[0125]

[0126] In the formula, The sample covariance matrix; The number of samples for key style features;

[0127] For covariance matrix Perform eigenvalue decomposition and select the eigenvectors corresponding to the top 512 largest eigenvalues ​​to form a dimension reduction matrix;

[0128] The lightweight feature vector is obtained by multiplying the key style feature vector by the dimensionality reduction matrix, expressed by the following formula:

[0129]

[0130] In the formula, For lightweight feature vectors, For key style feature vectors, This is a dimension-reduced matrix.

[0131] In this embodiment, the feature dimension was reduced from 2048 to 512 after dimensionality reduction, the amount of data was reduced by 75%, and the time for feature extraction of a single model was reduced from 28.6 seconds to 3.2 seconds, improving efficiency by 8 times.

[0132] The lightweight features and copyright information are vector-concatenated to obtain the copyright concatenation features. A zero-watermark sequence of the copyright concatenation features is then generated using hash operations, specifically including:

[0133]

[0134] In the formula, To construct feature vectors for copyright, For lightweight feature vectors, For identification coding, Numerical encoding for timestamps. Encoding for the lung model;

[0135] Copyright splicing feature vector conduct The zero-watermark sequence obtained by hash operation is expressed by the following formula:

[0136]

[0137] In the formula, It is a zero-watermark sequence. for Hash operation function.

[0138] like Figure 2 As shown, the zero-watermark verification process includes: calculating the original zero-watermark sequence. With the sequence to be verified of The value is used to evaluate feature similarity, and the formula is:

[0139]

[0140] in, It is the original zero-watermark sequence and the sequence to be verified The order of positions, , They are respectively , The A binary value (0 or 1).

[0141] Anatomical semantic secondary verification: Validation Semantic similarity between features of each key subregion (lung lobe, lesion) and corresponding subregion features in the original model (The calculation method is the same as step (1)), and it is necessary to satisfy all lung lobe subregions. ≥0.75, lesion subregion ≥0.8;

[0142] Copyright Determination Rules: If If the value is ≥0.85 and the semantic analysis passes, the sequence to be verified is determined. Copyright valid; if If the value is less than 0.85 or semantic verification fails, the sequence to be verified is determined. There is infringement (e.g., partial alteration would lead to...). (Asymmetric pruning can lead to semantic validation failures). Under a 20% asymmetric pruning attack, the method's... The value reached 0.85, which is 37% higher than the global feature zero watermarking technology.

[0143] Example 2

[0144] This embodiment provides an electronic terminal, characterized in that it includes a processor and a storage medium; the storage medium is used to store instructions; the processor is used to perform operations according to the instructions to execute the steps of the zero watermark generation method described in Embodiment 1.

[0145] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0146] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0147] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0148] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0149] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for generating a zero-watermark lightweight three-dimensional lung model, characterized in that, include: Adaptive sparse voxel sampling is performed on the original 3D lung model to obtain a simplified 3D lung model; The 3D U-Net network was used to segment the anatomical subregions of a simplified 3D lung model and extract subregion style features. Calculate the cosine similarity between sub-region style features and anatomical semantic labels. Based on cosine similarity The sub-region style features are verified and filtered to obtain independent sub-feature vectors; Multi-scale features are obtained by hierarchically extracting independent sub-feature vectors; global style features are obtained by cross-scale fusion of multi-scale features. Global style features are input into the 3D MobileViT lightweight network to obtain key style features, and PCA feature dimensionality reduction is performed on the key style features to obtain lightweight features. The lightweight features and copyright information are vector-concatenated to obtain the copyright concatenation features. A zero-watermark sequence of the copyright concatenation features is then generated using hash operations, specifically including: ； In the formula, To construct feature vectors for copyright, For lightweight feature vectors, For identification coding, Numerical encoding for timestamps. Encoding for the lung model; Copyright splicing feature vector conduct The zero-watermark sequence obtained by hash operation is expressed by the following formula: ； In the formula, It is a zero-watermark sequence. for Hash operation function.

2. The zero-watermark generation method according to claim 1, characterized in that, Adaptive sparse voxel sampling is performed on the original 3D lung model to obtain a simplified 3D lung model, specifically including: Calculate the voxel gray variance of the original three-dimensional lung model, divide the original three-dimensional lung model into lung voxel dense region and lung voxel sparse region according to the voxel gray variance, and set the sampling rate of lung voxel dense region and lung voxel sparse region. The first voxel sampling feature is obtained by randomly sampling the sparse voxel region of the lung according to the set sampling rate, and the second voxel sampling feature is obtained by grid sampling the dense voxel region of the lung according to the set sampling rate. A simplified three-dimensional lung model was obtained by constructing the first voxel sampling features and the second voxel sampling features.

3. The zero-watermark generation method according to claim 1, characterized in that, The 3D U-Net network was used to segment anatomical subregions and extract subregion style features from a simplified 3D lung model. Specifically, this included: The 3D U-Net network includes an input layer, an encoder, a decoder, and an output layer connected in sequence; the encoder contains four 3D convolutional modules. The decoder contains four layers of deconvolution modules. The encoder contains a 4-layer 3D convolutional module. With the deconvolution module in the decoder Through residual connections; where... This is the sequence number of the 3D convolution module. This is the sequence number of the deconvolution module. ; The simplified 3D lung model is input into the 3D U-Net network through the input layer. The encoder upsamples the simplified 3D lung model to obtain 3D spatial information, and the decoder downsamples the 3D spatial information. The output layer outputs the sub-region segmentation features and sub-region types. Based on the subregion type, lung lobe subregion features, tracheal tree subregion features, and lesion subregion features are extracted from the subregion segmentation features, respectively. The features of the lung lobe subregion include the mean gray value of the lung lobe, the variance of the lung lobe gray value, the skewness of the lung lobe gray value distribution, and the kurtosis of the lung lobe gray value distribution; the features of the tracheal tree subregion include the number of tracheal branches, the angle between tracheal branches, and the length of tracheal branches; the features of the lesion subregion include the mean curvature of the lesion surface and the Gaussian curvature of the lesion surface. The features of the lung lobe subregion, tracheal tree subregion, and lesion subregion are used as subregion style features.

4. The zero-watermark generation method according to claim 1, characterized in that, Based on cosine similarity The sub-region style features are validated and filtered to obtain independent sub-feature vectors, specifically including: If cosine similarity <Set a similarity threshold A to determine sub-region style feature extraction anomalies, and re-use the 3D U-Net network to perform anatomical sub-region segmentation and extract sub-region style features of the simplified 3D lung model; If cosine similarity ≥ Set a similarity threshold A, and generate independent sub-feature vectors from the style features of the sub-region.

5. The zero-watermark generation method according to claim 1, characterized in that, Multi-scale features are obtained by hierarchically extracting independent sub-feature vectors, specifically including: Voxel gray-level co-occurrence matrix features are extracted from independent sub-feature vectors according to a set microscale. These voxel gray-level co-occurrence matrix features include contrast. Correlation and energy Ene; Based on the set mesoscale, the topological features of bronchial branches and the spatial distribution features of lung lobes are extracted from the independent sub-feature vectors. Based on the set macroscopic scale, the spatial location features and morphological features of the lung lobes are extracted from the independent sub-feature vectors; the spatial location features of the lung lobes are the coordinates of the lung lobe center and the proportion of the lung lobe volume; the morphological features of the lung lobes are the principal axis ratio of the overall lung shape.

6. The zero-watermark generation method according to claim 5, characterized in that, Cross-scale fusion of multi-scale features yields global style features, specifically including: Z-score standardization was performed on the voxel gray-level co-occurrence matrix features, bronchial branch topological features, pulmonary lobule spatial distribution features, pulmonary lobe spatial location features, and pulmonary lobe morphological features to obtain micro-scale feature vectors, meso-scale feature vectors, and macro-scale feature vectors, respectively. Attention weights for microscale, mesoscale, and macroscale eigenvectors are calculated using linear transformations and the Softmax function, expressed as follows: ； ； In the formula, For attention weights, For scale type index identifier; It is an exponential function; It is a linear transformation function of the scale eigenvectors; It is a standardized feature vector with a defined scale; For microscale indexing; For mesoscale indexing; For macroscopic scale indexing; The weight matrix is ​​a linear transformation matrix; For bias terms; The global style feature vector is obtained by weighting the micro-scale feature vector, meso-scale feature vector, and macro-scale feature vector according to attention weights, and then summing them. The formula is as follows: ； In the formula, This is the global style feature vector; It is the attention weight at the microscale. It is a standardized vector at the microscale; It is the attention weight at the mesoscale. It is a standardized vector at the mesoscale; It is the attention weight at the macro scale. It is a standardized vector at a macroscopic scale.

7. The zero-watermark generation method according to claim 1, characterized in that, The 3D MobileViT lightweight network includes an input layer, three depthwise separable convolutional blocks, two local Transformer attention blocks, and an output layer connected in sequence. The depthwise separable convolutional blocks include 3D depthwise convolution, ReLU6 activation function, 3D pointwise convolution, and normalization processing layer in sequence.

8. The zero-watermark generation method according to claim 1, characterized in that, PCA feature reduction is performed on key style features to obtain lightweight features, specifically including: Calculate the mean vector of key style features The formula is as follows: ； In the formula, This is the mean vector of the key style features. The number of samples for key style features; It is the index of the key style feature vector; For the first Key style feature vectors; Based on the mean vector of key style features and the Key style feature vectors The formula for calculating the sample covariance matrix is ​​as follows: ； In the formula, The sample covariance matrix; The number of samples for key style features; For covariance matrix Perform eigenvalue decomposition and select the eigenvectors corresponding to the top 512 largest eigenvalues ​​to form a dimension reduction matrix; The lightweight feature vector is obtained by multiplying the key style feature vector by the dimensionality reduction matrix, expressed by the following formula: ； In the formula, For lightweight feature vectors, For key style feature vectors, This is a dimension-reduced matrix.

9. An electronic terminal, characterized in that, It includes a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the steps of the zero watermark generation method according to any one of claims 1 to 8.

Images